debug: add per-step timing logs to identify hang vs slow compute

Logs step_ms for first 5 steps and every 50th step thereafter.
If steps aren't completing, the graph is hanging.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -1505,7 +1505,15 @@ impl DQNTrainer {
                 let _fused_result = fused.run_full_step(&batch, &mut *agent, &self.device)
                     .context("Fused CUDA training step failed")?;
 
-                fused_total_us += fused_start.elapsed().as_micros() as u64;
+                let step_us = fused_start.elapsed().as_micros() as u64;
+                fused_total_us += step_us;
+                if train_step_count < 5 || train_step_count % 50 == 0 {
+                    tracing::info!(
+                        step = train_step_count,
+                        step_ms = step_us / 1000,
+                        "STEP_TIMING"
+                    );
+                }
 
                 let guard_start = std::time::Instant::now();
                 // Run guard every 5th step — NaN/loss checks add ~12ms/step overhead.